Subjet : Facial recognition for biometric authentication
Biometrics is present in our daily lives with digital passport, access control or even video surveillance applications. Among all the biometric modalities, facial recognition is a challenge. Indeed, facial recognition is a non-contact technique and is possible without the consent of the individual. However, many problems still exist, such as variations in illumination or facial orientation during acquisition. In addition, an individual changes over time (haircut, glasses, age…).
The purpose of this master thesis was firstly to make a bibliography of existing biometric methods, whether facial recognition, iris recognition or a fingerprint. We have seen that the conventional approaches in face recognition is based on the reduction of information using methods such as Principal Component Analysis (PCA) or more sophisticated methods such as fisherface. A bibliography of learning methods was also performed. Learning methods such as K-nearest neighbors, neural networks or support vector machines were studied.
In a second step, a biometric recognition algorithm using invariant descriptors and support vector machines was proposed. This recognition method is an adaptation to the face of the method proposed by Anant Choksuriwong in his PhD thesis entitled Interpretation scenes for robotics for object recognition. Invariant descriptors are conventionally used in object recognition. The fact that these descriptors are invariant is then particularly interesting to deal with the problems of illumination and orientation, and thus adapts well to the face recognition. A comparative study of several descriptor invariants, such as Hu moments, Fourier-Mellin descriptors or Zernike moments, gives results equivalent to the best results of the state art. Indeed, using the AR face database, containing faces of 120 individuals, we achieved a good recognition rate up to 97.48% with Zernike moment.